AWS Community Day CEE is tomorrow. We will be there — and here is what we are listening for.

Why we are in the room

The post we published on 11 September starts from a decision, not a box. Agents get built in the cloud, because that is where the developers are; part of the inference has to stay on the customer's own floor, because that is where the data is. What actually has to be solved is which model answers a given request and which endpoint serves it. So we join our GPU nodes to the customer's EKS cluster as Amazon EKS Hybrid Nodes — our hardware, an AWS control plane — and put our AI Gateway in front of both sides, so the platform picks the model and the endpoint per use case, on sensitivity, cost, latency and data residency, with one governance and observability plane over both. It is offered as a fixed-price six-week pilot on a real customer case.

Three pieces of that architecture were unfinished when we published it. Three sessions on Thursday are about exactly those three pieces, which is why they are the picks and not just the interesting talks.

• Routing. One Endpoint, Every Model is the session, and its own abstract makes our point better than we do: Bedrock now puts 19 model providers behind one API, yet its built-in Intelligent Prompt Routing only picks between two sibling models. Routing across providers — and out of the region to a self-hosted endpoint — is still the customer's problem to solve. The speaker solves it with a LiteLLM gateway on AWS parts. We solve it with our own gateway on-prem. We would like to see the seam from his side.
• Guardrails. Responsible AI in Practice is the policy layer, and the hard version of the question is ours: a guardrail configured in Bedrock has to hold when half the stack is not in AWS. If the policy only binds the calls that stay inside the region, it is not a policy, it is a default.
• Observability you are allowed to keep. Trust but Verify is about masking and tokenisation between the application and Bedrock — so the debug trace you keep is not itself the data leak. Everyone who has run a regulated workload knows the trade: the trace that would let you explain last Tuesday is the trace your DPO would rather you never wrote.

The anchor

The one session we would not move for anything is the ninety-minute hands-on workshop, Create Your Own Role-Playing Game with Agentic AI using Strands Agents. The game is the vehicle; Strands is the agent framework in our own stack. It is Apache-2.0. It ships first-party model providers for Amazon Bedrock, Ollama, LiteLLM and OpenAI, and its OpenAI provider takes a custom base_url — the documentation's words are "to connect to a custom OpenAI-compatible server" — which is precisely how an agent built on AWS points its inference at a vLLM endpoint on somebody else's GPU. That provider abstraction is the hinge the whole hybrid turns on. Ninety minutes with the people who wrote it is worth more than a week of reading the source.

One more reason, and it is a personal one

I come to this one straight off building a Microsoft Foundry landing zone for an enterprise team — the tenant, the networking, the identity, the guardrails, the whole floor an agent platform stands on. Colleagues from that team will be in the room on Thursday too. So a real purpose of the day is a comparison I cannot do from documentation: how Bedrock and agentic workflows carry enterprise workloads, measured against what I have just finished building on the other platform. Whatever the answer turns out to be, it goes back to those colleagues on Friday.

What we are listening for

We read the whole programme this morning. Nineteen sessions in one day across three rooms — sixteen talks and three ninety-minute workshops. Twelve of the nineteen carry the Artificial Intelligence & Machine Learning tag, and four of those twelve also carry Cost Optimization & FinOps.

It is an unusually practitioner-heavy agenda. Agents that have been in production for months rather than in a demo. Guardrails. Compliance. What the bill looks like. Very little of it is about what a model can do and most of it is about what happens to a team after the model is already there, which is rarer than it should be. The organisers deserve the credit for that.

The thing we will be listening for is the question sitting underneath almost every one of those AI sessions, which no session on the agenda asks out loud: where does the inference actually run, and whose invoice grows when it gets busy. The four cost sessions are that same question in its billing form — what it costs is downstream of where it runs. And the programme does have a Cloud Migration & Hybrid Architectures track: exactly one of the twelve AI sessions also sits in it, and that one is about getting AI projects finished with the partner ecosystem. A real problem, and a different one.

So the question is in the room all day and on the agenda nowhere. That is not a criticism of the event — it is an accurate picture of where the market is, and it is the gap we are building into.

The sessions we would not miss

Grouped by what we care about, quoted from the public programme, with the rooms as the organisers name them: Üvegterem is the main stage, Akvárium the breakout room, Hajó the workshops. Times are as published on 16 September; the organisers reserve the right to change the programme and the schedule, so check the agenda on the morning.

One day, no replays

The AI Summit had an event app with recordings, so a clash was recoverable. This one is a single day in three rooms and nothing on the site promises replays, which makes 14:00 a genuine choice: the gateway session in Akvárium or the infrastructure-as-context session in Üvegterem. We will take the gateway and read the other one's slides afterwards.

The question we will ask

One question, asked of everyone we talk to:

It is the same question as the one we asked at the AI Summit, moved one layer down the stack. The answer is always specific. Sometimes it is "in the region, and we have stopped looking at the bill". Sometimes it is "we do not know, the vendor handles it", which is the interesting answer. Everything we build exists so that a company can give a precise answer to both halves without hiring a platform team.

We will publish what we learn here afterwards, including the parts that contradict us.

Practical details

• When: Thursday, 17 September 2026. Doors 08:00, first keynote 09:05, closing keynote 16:10.
• Where: Budapest. Three rooms: Üvegterem (main stage), Akvárium (breakout), Hajó (workshops).
• Programme: awscommunity.eu. The organisers reserve the right to change the programme and the schedule, so the times above are the 16 September state and nothing more.
• Language: the sessions are in English.
• A note on AWS: Amazon Web Services and AWS are trademarks of Amazon.com, Inc. or its affiliates. ATAILA is not affiliated with, sponsored by or endorsed by AWS. The work described above is our own, on a customer case.